Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

N-BEATS-MOE: N-BEATS with a Mixture-of-Experts Layer for Heterogeneous Time Series Forecasting

T0 review · 3 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that replacing the fixed sum of block forecasts in N-BEATS with a gate-weighted mixture of experts yields consistent accuracy gains on heterogeneous multi-series benchmarks and adds an interpretable routing signal.

desk verdict A clean, incremental MoE-on-N-BEATS paper whose central accuracy claim is weakened by single-run point estimates and tiny margins; worth referee time mainly to force proper statistical validation. read the letter →

arxiv 2508.07490 v1 pith:QLKDJVF6 submitted 2025-08-10 cs.LG stat.ML

Ricardo Matos , Luis Roque , Vitor Cerqueira This is my paper
classification cs.LGstat.ML
keywords timeseriesforecastingN-BEATSmixtureofexpertsgatingnetworkheterogeneousforecastinterpretabilitySMAPEglobalmodels
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish a focused architectural improvement: in N-BEATS, block forecasts are normally added together, while N-BEATS-MOE multiplies each block's forecast by a learned gate weight before summing, with the weights coming from a linear layer plus softmax on the model input. The claim is that this lets the network emphasise the right basis (trend, seasonal, or identity) for each series, which should help when one global model must forecast thousands of heterogeneous series. Evaluated on 12 benchmark datasets from M1, Tourism, M3, and M4, the paper reports lower SMAPE than N-BEATS in 6 of 9 frequency–dataset cells for the mixed-domain competitions, while N-BEATS wins on the single-domain Tourism data and on M3 quarterly and M4 monthly. The same gate weights are offered as an interpretability device: the largest weight shows which expert dominates a given forecast. If the accuracy result holds up, this is a low-cost drop-in modification to a widely used forecasting architecture.

What carries the argument

The carrying mechanism is the mixture-of-experts aggregation layer, Eqs. (2)–(3): $\hat{y} = \sum_{\ell=1}^{B} \hat{G}_\ell \cdot \hat{y}_\ell$ with $\hat{G}_\ell = \operatorname{softmax}_\ell(\operatorname{LINEAR}_\ell(x_0))$, where $x_0$ is the LayerNorm-normalised model input. It replaces the fixed equal-weight sum in N-BEATS, so each block output is scaled before being combined; the softmax keeps the weights a convex combination, and the LayerNorm is introduced to prevent the gate from collapsing onto one expert. This mechanism carries the whole argument: it is the only change to the architecture, it is what lets the network route heterogeneous series to different basis functions, and it

What would settle it

Re-run N-BEATS and N-BEATS-MOE on the M1, M3, and M4 frequency groups with multiple seeds and compute paired bootstrap intervals over per-series SMAPE differences; if the intervals cross zero in most of the cells where N-BEATS-MOE is reported better, the improvement claim is indistinguishable from noise. A second check: freeze or randomise the gate weights and retrain only the rest of the network; if accuracy stays roughly the same, the benefit comes from the weighted-sum parameterisation rather than from input-dependent routing.

Watch

Extended reading notes

Core claim

N-BEATS-MOE is N-BEATS with the final aggregation step changed. Instead of $\hat{y} = \sum_\ell \hat{y}_\ell$, the model computes $\hat{y} = \sum_\ell \hat{G}_\ell \hat{y}_\ell$, where $\hat{G}_\ell = \operatorname{softmax}_\ell(\operatorname{LINEAR}_\ell(x_0))$ over a LayerNorm-normalised input, so each block—the trend stack, seasonal stack, or identity stack—acts as an expert weighted by its relevance to the current series. The paper's central discovery claim is that this gate weighting gives consistent improvements over N-BEATS on datasets composed of heterogeneous time series: across M1, M3, and M4, the proposed model beats N-BEATS in 6 of 9 frequency–dataset cells, while on the single-d

Load-bearing premise

The load-bearing premise is that each SMAPE value in Table 3 is a stable measurement, so differences as small as 0.02 points (M3 quarterly: 9.01 for N-BEATS versus 9.03 for N-BEATS-MOE) are meaningful; the paper reports no repeated seeds, confidence intervals, or statistical tests to support that premise.

Editorial extensions

If this is right

  • On the mixed-domain benchmark families M1, M3, and M4, N-BEATS-MOE reports lower SMAPE than N-BEATS in six of the nine frequency–dataset cells, so the claimed gain concentrates where series come from varied domains.
  • On the single-domain Tourism family, N-BEATS-MOE is worse in all three frequency groups, so the paper's own results bound the method's benefit to heterogeneous settings.
  • The gate weights are claimed to add interpretability: the softmax values identify the most relevant expert per input, and the decomposition analysis shows the gate can down-weight an over-amplified trend component and improve seasonal tracking.
  • The paper reports that most MoE variants—MoEBlock, MoEShared, and MoEScaled—also outperform plain N-BEATS on the same benchmarks, indicating the gating idea is not tied to one specific wiring of the layer.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural control the paper does not run would replace the learned gate with fixed or random weights of the same shape; if SMAPE gains persist, the improvement is a parameterisation effect rather than evidence for input-dependent routing.
  • The gate is a function of the raw input only, so the same weights could be reused as a per-series fingerprint for clustering, anomaly detection, or measuring distribution shift across datasets—uses the paper leaves unexplored.
  • The Tourism results predict that gate-based gains should shrink as series homogeneity grows; this can be tested directly by subsampling a mixed dataset like M3 into more homogeneous clusters and checking whether N-BEATS-MOE's advantage over N-BEATS fades.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes N-BEATS-MOE, a modification of N-BEATS in which the sum of block outputs is replaced by a softmax-gated weighted sum (Eqs. 1–3). A linear gate applied to the LayerNorm-normalized input produces a weight for each block, and the final forecast is the weighted sum of the per-block forecasts. The authors evaluate the method on 12 dataset-frequency groups from M1, Tourism, M3, and M4, reporting SMAPE. They claim consistent improvements over N-BEATS, particularly on heterogeneous datasets, and further claim that the learned gate weights add interpretability by revealing which expert (trend, seasonal, etc.) is most relevant for a given series. An STL-decomposition analysis (Table 4) is presented to examine specialization, with the paper itself acknowledging inconclusive alignment.

Significance. If the accuracy claim were rigorously established, N-BEATS-MOE would be a modest but useful contribution: it is a minimal, clearly specified architectural change that can be dropped into an existing N-BEATS pipeline, and the gating mechanism is a plausible way to let a global model adapt to heterogeneous series. The interpretability angle is also attractive, though the paper's own results are mixed. The manuscript does not provide code, multiple seeds, confidence intervals, or statistical tests, however, so the central empirical claim is not yet supported at the standard expected in this area. The architecture description (Eqs. 1–3) is clear, and the side-by-side comparison against N-BEATS on standard benchmarks is a useful starting point, but the evidence base is currently too thin for a firm positive verdict.

major comments (3)
  1. [Section 4.4 / Table 3] The central claim of 'consistent improvements over N-BEATS' rests entirely on single-point SMAPE estimates. The differences are frequently within optimization noise: M3 quarterly is 9.01 (N-BEATS) vs 9.03 (N-BEATS-MOE), and the method loses on the largest dataset group, M4 monthly (12.79 vs 12.99), as well as on all three Tourism groups. No seeds, confidence intervals, or paired per-series tests are reported. Since the models are global, the effective sample size is the number of series, so a paired comparison on the test series would be both feasible and appropriate. Without this, the 6-of-9 win count does not support the abstract's 'consistent improvements' or Section 4.4's 'consistent improvements over N-BEATS, especially in datasets composed of heterogeneous time series.' The claim needs to be re-evaluated with at least 5 seeds and reported mean±std, plus a significance test or effec
  2. [Sections 4.2–4.3] The manuscript as provided jumps from Section 4.1 directly to Section 4.4; the experimental setup, hyperparameter search, and model configuration (number of stacks, blocks, basis degree d, learning rate, batch size, epochs, loss, and any regularization) are absent. This is load-bearing because the comparison's fairness cannot be assessed: N-BEATS-MOE adds a gate with additional parameters, and the paper does not state whether the base N-BEATS hyperparameters were tuned separately or reused. Code is also not provided (only a footnote link to an external MoE repository). The missing setup section is a serious omission that must be supplied for the results to be reproducible and for the comparison to be meaningful.
  3. [Section 4.5 / Table 4 / Figure 5] The interpretability analysis does not establish that gate weights correspond to meaningful expert specialization. Table 4 reports selection ratios on only the monthly M1, M3, and M4 datasets, and the paper admits the alignment with STL components is 'inconclusive' (Section 5). Figure 5 is a cherry-picked set of four series hand-picked because N-BEATS-MOE did better; no systematic or quantitative evaluation is provided for the claim that high gate weight implies qualitative contribution. The interpretability claim in the abstract and Section 1 should be softened to a hypothesis, or supported with a more rigorous analysis (e.g., measuring correlation between gate weights and series characteristics over the full test set, or ablating the gate to show that the weighted combination yields different behavior than the unweighted sum).
minor comments (6)
  1. [Section 4.4] Typo: 'Turism' should be 'Tourism' in the sentence 'However, in the Turism datasets...'.
  2. [Section 5] Typo: 'the results where inconclusive' should be 'were inconclusive'.
  3. [Section 4.1] The SMAPE formula is referenced as 'defined as follows' but the actual equation is not present in the provided text; please ensure the formula appears in the final version.
  4. [Figure 5] The subplot labels and the gate-weight annotations are hard to read in the current rendering; larger fonts and clear annotations are needed.
  5. [Eq. (3) / Section 3] The notation 'LINEAR_l' is used per block, but the text says 'a learnable affine transformation' (singular). Clarify whether each block has its own linear layer or whether this is a shared linear layer followed by a per-block index; this matters for parameter count and implementation.
  6. [General] The method is called Mixture-of-Experts but the gate is a dense softmax combination, not a sparse expert selection. Readers may expect sparsity; a sentence clarifying that this design is a 'soft MoE' or 'dense gating' would avoid confusion.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the gate is a learned parameter and accuracy is tested on external benchmarks; the main weakness is statistical, not circular.

full rationale

The paper's core architectural claim is that replacing the N-BEATS summation with a learned gate-weighted sum improves forecasting on heterogeneous datasets. The relevant equations are Eq. (2): y_hat = sum_l G_l * y_hat_l, and Eq. (3): G_l = softmax_l(LINEAR_l(x0)). The gate weights are not defined in terms of the evaluation metric or the final forecast; they are learned parameters in a standard supervised training loop. The reported SMAPE values come from held-out test sets of external competition benchmarks (M1, M3, M4, Tourism), so the accuracy claim is empirically falsifiable and not derived from the model's own assumptions. The N-BEATS backbone, including the trend/seasonal basis expansions, is cited from the original external paper [22]. There is no load-bearing self-citation: the only author co-authored reference (Cerqueira et al. [2]) appears in the bibliography without an in-text argumentative use. The interpretability claim (gate weights indicate expert relevance) is explicitly hedged in Section 4.5, where the authors report that the STL-decomposition alignment is 'inconclusive' and that gate behavior is 'sensitive to dataset-specific characteristics.' This honesty further reduces any concern that the authors are renaming a known result as a derivation. The legitimate criticism of this paper is statistical robustness: single-run point estimates with differences as small as 0.02 SMAPE (M3 quarterly: 9.01 vs 9.03) are treated as meaningful without seeds, confidence intervals, or significance tests. That is a correctness-risk issue, not a circularity issue, and the instructions are explicit that 'not standard consensus' and statistical fragility are not circularity arguments. Accordingly, no circular step can be identified, and the score is 0.

Assumptions & free parameters 2 free parameters · 4 assumptions · 1 invented entities

The central empirical claim depends on the learned gate parameters, on unspecified N-BEATS hyperparameters, and on domain assumptions about dataset heterogeneity and about what gate weights mean. No new physical entities are proposed.

free parameters (2)
  • Gating network weights and bias (LINEAR_l and softmax in Eq. 3) = not reported
    Learned on each benchmark; the accuracy of the gated forecast depends on these fitted parameters.
  • N-BEATS base hyperparameters (number of stacks, blocks, basis degree d, learning rate, batch size) = not reported
    Inherited from the original N-BEATS setup, but not specified in this manuscript; they influence both the baseline and N-BEATS-MOE results.
assumptions (4)
  • domain assumption SMAPE is the appropriate error metric for comparing forecasts on these competition benchmarks.
    The paper evaluates all models with SMAPE and does not report alternative metrics or distributional summaries.
  • domain assumption M1, M3 and M4 are heterogeneous because they contain series from mixed domains, while Tourism is homogeneous.
    Section 4.1 and 4.4 use this grouping to explain why N-BEATS-MOE wins more often on M1/M3/M4 and loses on Tourism.
  • ad hoc to paper The softmax gate weights indicate the relative importance of each block for a given series.
    Section 4.5 interprets gate weights as evidence of expert specialization, but the paper's own conclusion says the alignment with STL decomposition was inconclusive.
  • domain assumption LayerNorm applied to the model input helps prevent mode collapse in the gate.
    Section 3 states this without an experiment isolating the effect of LayerNorm.
invented entities (1)
  • MoE gating layer over N-BEATS blocks (LINEAR_l followed by softmax)
    purpose: Produces per-series weights that reweight the block forecasts in Eq. 2.
    This is a new architectural component with no external falsifiable handle beyond the benchmark numbers; its claimed interpretability value is not independently validated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of N-BEATS-MOE: N-BEATS with a Mixture-of-Experts Layer for Heterogeneous Time Series Forecasting." pith.science (2026). https://pith.science/paper/QLKDJVF6

@misc{pith2026250807490,
  author       = {Pith},
  title        = {Pith review of: N-BEATS-MOE: N-BEATS with a Mixture-of-Experts Layer for Heterogeneous Time Series Forecasting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QLKDJVF6}},
  note         = {Machine review of arXiv:2508.07490}
}
read the original abstract

Deep learning approaches are increasingly relevant for time series forecasting tasks. Methods such as N-BEATS, which is built on stacks of multilayer perceptrons (MLPs) blocks, have achieved state-of-the-art results on benchmark datasets and competitions. N-BEATS is also more interpretable relative to other deep learning approaches, as it decomposes forecasts into different time series components, such as trend and seasonality. In this work, we present N-BEATS-MOE, an extension of N-BEATS based on a Mixture-of-Experts (MoE) layer. N-BEATS-MOE employs a dynamic block weighting strategy based on a gating network which allows the model to better adapt to the characteristics of each time series. We also hypothesize that the gating mechanism provides additional interpretability by identifying which expert is most relevant for each series. We evaluate our method across 12 benchmark datasets against several approaches, achieving consistent improvements on several datasets, especially those composed of heterogeneous time series.

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Hopformer: Homogeneity-Pursuit Transformer for Time Series Forecasting

    stat.ML 2026-07 reject novelty 5.0 of 10

    A two-stage forecaster (SPA trend extraction + LoRA-fine-tuned residual Transformer) that the paper claims beats prior models by 6.56% MASE, though the claim is not robust to its own extended baseline tables.

Reference graph

Works this paper leans on

34 extracted references · 15 canonical work pages · cited by 1 Pith paper

  1. [1]

    International Journal of Forecasting27(3), 822–844 (Jul 2011)

    Athanasopoulos, G., Hyndman, R.J., Song, H., Wu, D.C.: The tourism forecast- ing competition. International Journal of Forecasting27(3), 822–844 (Jul 2011). https://doi.org/10.1016/j.ijforecast.2010.04.009, https://linkinghub.elsevier. com/retrieve/pii/S016920701000107X

    work page doi:10.1016/j.ijforecast.2010.04.009 2011

  2. [2]

    Machine Learning 108, 913–944 (2019)

    Cerqueira, V., Torgo, L., Pinto, F., Soares, C.: Arbitrage of forecasting experts. Machine Learning 108, 913–944 (2019)

    2019

  3. [3]

    https://doi.org/10.48550/arXiv.2201.12886, http://arxiv.org/ abs/2201.12886, arXiv:2201.12886 [cs]

    Challu, C., Olivares, K.G., Oreshkin, B.N., Garza, F., Mergenthaler-Canseco, M., Dubrawski,A.:N-HiTS:NeuralHierarchicalInterpolationforTimeSeriesForecast- ing (Nov 2022). https://doi.org/10.48550/arXiv.2201.12886, http://arxiv.org/ abs/2201.12886, arXiv:2201.12886 [cs]

    arXiv doi:10.48550/arxiv.2201.12886 2022

  4. [4]

    Cleveland, R.B., Cleveland, W.S., McRae, J.E., Terpenning, I., et al.: Stl: A seasonal-trend decomposition. J. off. Stat6(1), 3–73 (1990)

    work page 1990

  5. [5]

    https://doi.org/10.48550/arXiv.2401.06066, http://arxiv.org/abs/2401.06066, arXiv:2401.06066 [cs]

    Dai, D., Deng, C., Zhao, C., Xu, R.X., Gao, H., Chen, D., Li, J., Zeng, W., Yu, X., Wu, Y., Xie, Z., Li, Y.K., Huang, P., Luo, F., Ruan, C., Sui, Z., Liang, W.: DeepSeekMoE: Towards Ultimate Expert Specialization in Mixture-of- Experts Language Models (Jan 2024). https://doi.org/10.48550/arXiv.2401.06066, http://arxiv.org/abs/2401.06066, arXiv:2401.06066 [cs]

    arXiv doi:10.48550/arxiv.2401.06066 2024

  6. [6]

    arXiv preprint arXiv:2304.08424 (2023)

    Das, A., Kong, W., Leach, A., Mathur, S., Sen, R., Yu, R.: Long-term forecasting with tide: Time-series dense encoder. arXiv preprint arXiv:2304.08424 (2023)

    arXiv 2023

  7. [7]

    In: International conference on machine learning

    Du, N., Huang, Y., Dai, A.M., Tong, S., Lepikhin, D., Xu, Y., Krikun, M., Zhou, Y., Yu, A.W., Firat, O., et al.: Glam: Efficient scaling of language models with mixture-of-experts. In: International conference on machine learning. pp. 5547–

  8. [8]

    Han, X., Zhang, X., Wu, Y., Zhang, Z., Wu, Z.: Kan4tsf: Are kan and kan-based models effective for time series forecasting? arXiv preprint arXiv:2408.11306 (2024)

    arXiv 2024

Show all 34 references
  1. [9]

    OTexts (May 2018), google-Books-ID: _bBhDwAAQBAJ

    Hyndman, R.J., Athanasopoulos, G.: Forecasting: principles and practice. OTexts (May 2018), google-Books-ID: _bBhDwAAQBAJ

    2018

  2. [10]

    Neural Computation 3(1), 79–87 (Mar 1991)

    Jacobs, R.A., Jordan, M.I., Nowlan, S.J., Hinton, G.E.: Adaptive Mix- tures of Local Experts. Neural Computation 3(1), 79–87 (Mar 1991). https://doi.org/10.1162/neco.1991.3.1.79, https://ieeexplore.ieee.org/ document/6797059, conference Name: Neural Computation

    1991

  3. [11]

    International Journal of Forecasting 36(1), 167–177 (Jan 2020)

    Januschowski, T., Gasthaus, J., Wang, Y., Salinas, D., Flunkert, V., Bohlke-Schneider, M., Callot, L.: Criteria for classifying forecasting meth- ods. International Journal of Forecasting 36(1), 167–177 (Jan 2020). https://doi.org/10.1016/j.ijforecast.2019.05.008, https://www....

    2020 doi

  4. [12]

    Jiang, A.Q., Sablayrolles, A., Roux, A., Mensch, A., Savary, B., Bamford, C., Chaplot, D.S., Casas, D.d.l., Hanna, E.B., Bressand, F., Lengyel, G., Bour, G., Lample, G., Lavaud, L.R., Saulnier, L., Lachaux, M.A., Stock, P., Subramanian, S., Yang, S., Antoniak, S., Scao, T.L., ...

    2024 arXiv

  5. [13]

    International Journal of Forecasting 37(4), 1748–1764 (2021)

    Lim, B., Arık, S.Ö., Loeff, N., Pfister, T.: Temporal fusion transformers for inter- pretable multi-horizon time series forecasting. International Journal of Forecasting 37(4), 1748–1764 (2021)

    2021

  6. [14]

    https://doi.org/10.48550/arXiv.2410.10469, http://arxiv.org/abs/2410.10469, arXiv:2410.10469

    Liu, X., Liu, J., Woo, G., Aksu, T., Liang, Y., Zimmermann, R., Liu, C., Savarese, S., Xiong, C., Sahoo, D.: Moirai-MoE: Empowering Time Series Foundation Models with Sparse Mixture of Experts (Oct 2024). https://doi.org/10.48550/arXiv.2410.10469, http://arxiv.org/abs/2410.104...

    2024 doi arXiv

  7. [15]

    Journal of Forecasting 1(2), 111–153 (1982)

    Makridakis, S., Andersen, A., Carbone, R., Fildes, R., Hibon, M., Lewandowski, R., Newton, J., Parzen, E., Winkler, R.: The accuracy of extrapolation (time series) methods: Results of a forecasting competition. Journal of Forecasting 1(2), 111–153 (1982). https://doi.org/10.10...

    1982 doi

  8. [16]

    International Journal of Forecasting 16(4), 451–476 (Oct 2000)

    Makridakis, S., Hibon, M.: The M3-Competition: results, conclusions and implications. International Journal of Forecasting 16(4), 451–476 (Oct 2000). https://doi.org/10.1016/S0169-2070(00)00057-1, https://www.sciencedirect. com/science/article/pii/S0169207000000571

    doi

  9. [17]

    International Journal of Forecast- ing 34(4), 802–808 (Oct 2018)

    Makridakis, S., Spiliotis, E., Assimakopoulos, V.: The M4 Competition: Re- sults, findings, conclusion and way forward. International Journal of Forecast- ing 34(4), 802–808 (Oct 2018). https://doi.org/10.1016/j.ijforecast.2018.06.001, https://www.sciencedirect.com/science/art...

    2018 doi

  10. [18]

    International Journal of Forecasting38(4), 1346– 1364 (2022)

    Makridakis, S., Spiliotis, E., Assimakopoulos, V.: M5 accuracy competition: Re- sults, findings, and conclusions. International Journal of Forecasting38(4), 1346– 1364 (2022)

    2022

  11. [19]

    Artificial Intelligence Review 42, 275–293 (2014)

    Masoudnia, S., Ebrahimpour, R.: Mixture of experts: a literature survey. Artificial Intelligence Review 42, 275–293 (2014)

    2014

  12. [20]

    International Journal of Forecast- ing 37(4), 1632–1653 (Oct 2021)

    Montero-Manso, P., Hyndman, R.J.: Principles and algorithms for forecasting groups of time series: Locality and globality. International Journal of Forecast- ing 37(4), 1632–1653 (Oct 2021). https://doi.org/10.1016/j.ijforecast.2021.03.004, https://linkinghub.elsevier.com/retr...

    2021 doi

  13. [21]

    In: International Conference on Artificial Intelligence and Statistics

    Ni, R., Lin, Z., Wang, S., Fanti, G.: Mixture-of-Linear-Experts for Long-term Time Series Forecasting. In: International Conference on Artificial Intelligence and Statistics. pp. 4672–4680. PMLR (2024),https://proceedings.mlr.press/v238/ ni24a.html

    2024

  14. [22]

    https://doi.org/10.48550/arXiv.1905.10437, http://arxiv.org/abs/1905.10437, arXiv:1905.10437 N-BEATS with a Mixture-of-Experts Layer 15

    Oreshkin, B.N., Carpov, D., Chapados, N., Bengio, Y.: N-BEATS: Neural basis expansion analysis for interpretable time series forecasting (Feb 2020). https://doi.org/10.48550/arXiv.1905.10437, http://arxiv.org/abs/1905.10437, arXiv:1905.10437 N-BEATS with a Mixture-of-Experts Layer 15

    2020 doi arXiv

  15. [23]

    In: Proceedings of the 27th International Joint Conference on Artificial Intelligence

    Park, D.K., Yoo, S., Bahng, H., Choo, J., Park, N.: Megan: mixture of experts of generative adversarial networks for multimodal image generation. In: Proceedings of the 27th International Joint Conference on Artificial Intelligence. p. 878–884. IJCAI’18, AAAI Press (2018)

    2018

  16. [24]

    In: Proceedingsofthe35thInternationalConferenceonNeuralInformationProcessing Systems

    Riquelme, C., Puigcerver, J., Mustafa, B., Neumann, M., Jenatton, R., Pinto, A.S., Keysers, D., Houlsby, N.: Scaling vision with sparse mixture of experts. In: Proceedingsofthe35thInternationalConferenceonNeuralInformationProcessing Systems. NIPS ’21, Curran Associates Inc., R...

    2021

  17. [25]

    https://doi.org/10.48550/arXiv.1701.06538, http://arxiv

    Shazeer, N., Mirhoseini, A., Maziarz, K., Davis, A., Le, Q., Hinton, G., Dean, J.: Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer (Jan 2017). https://doi.org/10.48550/arXiv.1701.06538, http://arxiv. org/abs/1701.06538, arXiv:1701.06538 [cs]

    2017 doi arXiv

  18. [26]

    https://doi.org/10.48550/arXiv.2409.16040, http://arxiv.org/abs/2409.16040, arXiv:2409.16040 [cs]

    Shi, X., Wang, S., Nie, Y., Li, D., Ye, Z., Wen, Q., Jin, M.: Time-MoE: Billion- Scale Time Series Foundation Models with Mixture of Experts (Feb 2025). https://doi.org/10.48550/arXiv.2409.16040, http://arxiv.org/abs/2409.16040, arXiv:2409.16040 [cs]

    2025 doi arXiv

  19. [27]

    International journal of forecasting36(1), 75–85 (2020)

    Smyl, S.: A hybrid method of exponential smoothing and recurrent neural networks for time series forecasting. International journal of forecasting36(1), 75–85 (2020)

    2020

  20. [28]

    Sun, Y., Xie, Z., Eldele, E., Chen, D., Hu, Q., Wu, M.: Learning Pattern-Specific Experts for Time Series Forecasting Under Patch-level Distribution Shift (Oct 2024), http://arxiv.org/abs/2410.09836, arXiv:2410.09836 [cs]

    2024

  21. [29]

    box-jenkins methodology

    Tang, Z., De Almeida, C., Fishwick, P.A.: Time series forecasting using neural networks vs. box-jenkins methodology. Simulation57(5), 303–310 (1991)

    1991

  22. [30]

    Wang, G., Chen, Y., Gao, M., Wu, Z., Tang, J., Zhao, J.: A Time Series is Worth Five Experts: Heterogeneous Mixture of Experts for Traffic Flow Prediction (Sep 2024), http://arxiv.org/abs/2409.17440, arXiv:2409.17440

    2024 arXiv

  23. [31]

    https://doi.org/10.48550/arXiv.2408.15664, http://arxiv.org/abs/2408.15664, arXiv:2408.15664 [cs]

    Wang, L., Gao, H., Zhao, C., Sun, X., Dai, D.: Auxiliary-Loss- Free Load Balancing Strategy for Mixture-of-Experts (Aug 2024). https://doi.org/10.48550/arXiv.2408.15664, http://arxiv.org/abs/2408.15664, arXiv:2408.15664 [cs]

    2024 doi arXiv

  24. [32]

    Wu, H., Hu, T., Liu, Y., Zhou, H., Wang, J., Long, M.: Timesnet: Temporal 2d- variationmodelingforgeneraltimeseriesanalysis.arXivpreprintarXiv:2210.02186 (2022)

    2022 arXiv

  25. [33]

    In: 2023 IEEE/CVF International Conference on Com- puter Vision (ICCV)

    Zhang, Y., Cai, R., Chen, T., Zhang, G., Zhang, H., Chen, P.Y., Chang, S., Wang, Z., Liu, S.: Robust Mixture-of-Expert Training for Convolutional Neural Networks . In: 2023 IEEE/CVF International Conference on Com- puter Vision (ICCV). pp. 90–101. IEEE Computer Society, Los Al...

    2023

  26. [34]

    Zhou, Y., Lei, T., Liu, H., Du, N., Huang, Y., Zhao, V., Dai, A.M., Chen, Z., Le, Q.V., Laudon, J.: Mixture-of-Experts with Expert Choice Routing. Advances in Neural Information Processing Systems 35, 7103–7114 (Dec 2022), https://proceedings.neurips.cc/paper_files/paper/2022/...

    2022

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.